|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
base_model: |
|
- meta-llama/Llama-3.1-8B |
|
--- |
|
|
|
# Llama Scope |
|
|
|
[**Technical Report Link**](https://arxiv.org/abs/2410.20526) |
|
|
|
[**Use with OpenMOSS lm_sae Github Repo**](https://github.com/OpenMOSS/Language-Model-SAEs/blob/main/examples/loading_llamascope_saes.ipynb) |
|
|
|
[**Use with SAELens**] |
|
|
|
[**Explore in Neuronpedia**] |
|
|
|
Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models, yet scalable training remains a significant challenge. We introduce a suite of 256 improved TopK SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. |
|
|
|
This is a frontpage of all Llama Scope SAEs. Please see the following link for checkpoints. |
|
|
|
## Naming Convention |
|
|
|
L[Layer][Position]-[Expansion]x |
|
|
|
For instance, an SAE with 8x the hidden size of Llama-3.1-8B, i.e. 32K features, trained on the 15th post-MLP residual stream is called L15R-8x. |
|
|
|
## Checkpoints |
|
|
|
[**Llama-3.1-8B-LXR-8x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXR-8x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXA-8x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXA-8x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXM-8x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXM-8x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXTC-8x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXTC-8x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXR-32x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXR-32x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXA-32x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXA-32x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXM-32x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXM-32x/tree/main) |
|
|
|
[**Llama-3.1-8B-LXTC-32x**](https://huggingface.co/fnlp/Llama3_1-8B-Base-LXTC-32x/tree/main) |
|
|
|
## Llama Scope SAE Overview |
|
|
|
<center> |
|
|
|
| | **Llama Scope** | **Scaling Monosemanticity** | **GPT-4 SAE** | **Gemma Scope** | |
|
|-----------------------|:-----------------------------:|:------------------------------:|:--------------------------------:|:---------------------------------:| |
|
| **Models** | Llama-3.1 8B (Open Source) | Claude-3.0 Sonnet (Proprietary) | GPT-4 (Proprietary) | Gemma-2 2B & 9B (Open Source) | |
|
| **SAE Training Data** | SlimPajama | Proprietary | Proprietary | Proprietary, Sampled from Mesnard et al. (2024) | |
|
| **SAE Position (Layer)** | Every Layer | The Middle Layer | 5/6 Late Layer | Every Layer | |
|
| **SAE Position (Site)** | R, A, M, TC | R | R | R, A, M, TC | |
|
| **SAE Width (# Features)** | 32K, 128K | 1M, 4M, 34M | 128K, 1M, 16M | 16K, 64K, 128K, 256K - 1M (Partial) | |
|
| **SAE Width (Expansion Factor)** | 8x, 32x | Proprietary | Proprietary | 4.6x, 7.1x, 28.5x, 36.6x | |
|
| **Activation Function** | TopK-ReLU | ReLU | TopK-ReLU | JumpReLU | |
|
|
|
</center> |
|
|
|
|
|
## Citation |
|
|
|
Please cite as: |
|
|
|
``` |
|
@article{he2024llamascope, |
|
title={Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders}, |
|
author={He, Zhengfu and Shu, Wentao and Ge, Xuyang and Chen, Lingjie and Wang, Junxuan and Zhou, Yunhua and Liu, Frances and Guo, Qipeng and Huang, Xuanjing and Wu, Zuxuan and others}, |
|
journal={arXiv preprint arXiv:2410.20526}, |
|
year={2024} |
|
} |
|
``` |
|
|
